Actually, in a naive implementation it is very computationally intensive, especially with large convolution cores (like on iOS 7 where it appears huge) and high resolutions.

Even optimized implementations on GPU are still computationally intensive compared to alpha blending.